1 /*
2 * CDDL HEADER START
3 *
4 * The contents of this file are subject to the terms of the
5 * Common Development and Distribution License (the "License").
6 * You may not use this file except in compliance with the License.
7 *
8 * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9 * or http://www.opensolaris.org/os/licensing.
10 * See the License for the specific language governing permissions
11 * and limitations under the License.
12 *
13 * When distributing Covered Code, include this CDDL HEADER in each
14 * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15 * If applicable, add the following below this CDDL HEADER, with the
16 * fields enclosed by brackets "[]" replaced with your own identifying
17 * information: Portions Copyright [yyyy] [name of copyright owner]
18 *
19 * CDDL HEADER END
20 */
21 /*
22 * Copyright (c) 2005, 2010, Oracle and/or its affiliates. All rights reserved.
23 * Copyright (c) 2012, 2017 by Delphix. All rights reserved.
24 * Copyright (c) 2014 Spectra Logic Corporation, All rights reserved.
25 * Copyright (c) 2014 Integros [integros.com]
26 * Copyright 2017 RackTop Systems.
27 */
28
29 #include <sys/zfs_context.h>
30 #include <sys/dbuf.h>
31 #include <sys/dnode.h>
32 #include <sys/dmu.h>
33 #include <sys/dmu_impl.h>
34 #include <sys/dmu_tx.h>
35 #include <sys/dmu_objset.h>
36 #include <sys/dsl_dir.h>
37 #include <sys/dsl_dataset.h>
38 #include <sys/spa.h>
39 #include <sys/zio.h>
40 #include <sys/dmu_zfetch.h>
41 #include <sys/range_tree.h>
42
43 static kmem_cache_t *dnode_cache;
44 /*
45 * Define DNODE_STATS to turn on statistic gathering. By default, it is only
46 * turned on when DEBUG is also defined.
47 */
48 #ifdef DEBUG
49 #define DNODE_STATS
50 #endif /* DEBUG */
51
52 #ifdef DNODE_STATS
53 #define DNODE_STAT_ADD(stat) ((stat)++)
54 #else
55 #define DNODE_STAT_ADD(stat) /* nothing */
56 #endif /* DNODE_STATS */
57
58 static dnode_phys_t dnode_phys_zero;
59
60 int zfs_default_bs = SPA_MINBLOCKSHIFT;
61 int zfs_default_ibs = DN_MAX_INDBLKSHIFT;
62
63 SYSCTL_DECL(_vfs_zfs);
64 SYSCTL_INT(_vfs_zfs, OID_AUTO, default_bs, CTLFLAG_RWTUN,
65 &zfs_default_bs, 0, "Default dnode block shift");
66 SYSCTL_INT(_vfs_zfs, OID_AUTO, default_ibs, CTLFLAG_RWTUN,
67 &zfs_default_ibs, 0, "Default dnode indirect block shift");
68
69 #ifdef illumos
70 #ifdef _KERNEL
71 static kmem_cbrc_t dnode_move(void *, void *, size_t, void *);
72 #endif /* _KERNEL */
73 #endif
74
75 static int
dbuf_compare(const void * x1,const void * x2)76 dbuf_compare(const void *x1, const void *x2)
77 {
78 const dmu_buf_impl_t *d1 = x1;
79 const dmu_buf_impl_t *d2 = x2;
80
81 int cmp = AVL_CMP(d1->db_level, d2->db_level);
82 if (likely(cmp))
83 return (cmp);
84
85 cmp = AVL_CMP(d1->db_blkid, d2->db_blkid);
86 if (likely(cmp))
87 return (cmp);
88
89 if (d1->db_state == DB_SEARCH) {
90 ASSERT3S(d2->db_state, !=, DB_SEARCH);
91 return (-1);
92 } else if (d2->db_state == DB_SEARCH) {
93 ASSERT3S(d1->db_state, !=, DB_SEARCH);
94 return (1);
95 }
96
97 return (AVL_PCMP(d1, d2));
98 }
99
100 /* ARGSUSED */
101 static int
dnode_cons(void * arg,void * unused,int kmflag)102 dnode_cons(void *arg, void *unused, int kmflag)
103 {
104 dnode_t *dn = arg;
105 int i;
106
107 rw_init(&dn->dn_struct_rwlock, NULL, RW_DEFAULT, NULL);
108 mutex_init(&dn->dn_mtx, NULL, MUTEX_DEFAULT, NULL);
109 mutex_init(&dn->dn_dbufs_mtx, NULL, MUTEX_DEFAULT, NULL);
110 cv_init(&dn->dn_notxholds, NULL, CV_DEFAULT, NULL);
111
112 /*
113 * Every dbuf has a reference, and dropping a tracked reference is
114 * O(number of references), so don't track dn_holds.
115 */
116 refcount_create_untracked(&dn->dn_holds);
117 refcount_create(&dn->dn_tx_holds);
118 list_link_init(&dn->dn_link);
119
120 bzero(&dn->dn_next_nblkptr[0], sizeof (dn->dn_next_nblkptr));
121 bzero(&dn->dn_next_nlevels[0], sizeof (dn->dn_next_nlevels));
122 bzero(&dn->dn_next_indblkshift[0], sizeof (dn->dn_next_indblkshift));
123 bzero(&dn->dn_next_bonustype[0], sizeof (dn->dn_next_bonustype));
124 bzero(&dn->dn_rm_spillblk[0], sizeof (dn->dn_rm_spillblk));
125 bzero(&dn->dn_next_bonuslen[0], sizeof (dn->dn_next_bonuslen));
126 bzero(&dn->dn_next_blksz[0], sizeof (dn->dn_next_blksz));
127
128 for (i = 0; i < TXG_SIZE; i++) {
129 list_link_init(&dn->dn_dirty_link[i]);
130 dn->dn_free_ranges[i] = NULL;
131 list_create(&dn->dn_dirty_records[i],
132 sizeof (dbuf_dirty_record_t),
133 offsetof(dbuf_dirty_record_t, dr_dirty_node));
134 }
135
136 dn->dn_allocated_txg = 0;
137 dn->dn_free_txg = 0;
138 dn->dn_assigned_txg = 0;
139 dn->dn_dirtyctx = 0;
140 dn->dn_dirtyctx_firstset = NULL;
141 dn->dn_bonus = NULL;
142 dn->dn_have_spill = B_FALSE;
143 dn->dn_zio = NULL;
144 dn->dn_oldused = 0;
145 dn->dn_oldflags = 0;
146 dn->dn_olduid = 0;
147 dn->dn_oldgid = 0;
148 dn->dn_newuid = 0;
149 dn->dn_newgid = 0;
150 dn->dn_id_flags = 0;
151
152 dn->dn_dbufs_count = 0;
153 avl_create(&dn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
154 offsetof(dmu_buf_impl_t, db_link));
155
156 dn->dn_moved = 0;
157 POINTER_INVALIDATE(&dn->dn_objset);
158 return (0);
159 }
160
161 /* ARGSUSED */
162 static void
dnode_dest(void * arg,void * unused)163 dnode_dest(void *arg, void *unused)
164 {
165 int i;
166 dnode_t *dn = arg;
167
168 rw_destroy(&dn->dn_struct_rwlock);
169 mutex_destroy(&dn->dn_mtx);
170 mutex_destroy(&dn->dn_dbufs_mtx);
171 cv_destroy(&dn->dn_notxholds);
172 refcount_destroy(&dn->dn_holds);
173 refcount_destroy(&dn->dn_tx_holds);
174 ASSERT(!list_link_active(&dn->dn_link));
175
176 for (i = 0; i < TXG_SIZE; i++) {
177 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
178 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
179 list_destroy(&dn->dn_dirty_records[i]);
180 ASSERT0(dn->dn_next_nblkptr[i]);
181 ASSERT0(dn->dn_next_nlevels[i]);
182 ASSERT0(dn->dn_next_indblkshift[i]);
183 ASSERT0(dn->dn_next_bonustype[i]);
184 ASSERT0(dn->dn_rm_spillblk[i]);
185 ASSERT0(dn->dn_next_bonuslen[i]);
186 ASSERT0(dn->dn_next_blksz[i]);
187 }
188
189 ASSERT0(dn->dn_allocated_txg);
190 ASSERT0(dn->dn_free_txg);
191 ASSERT0(dn->dn_assigned_txg);
192 ASSERT0(dn->dn_dirtyctx);
193 ASSERT3P(dn->dn_dirtyctx_firstset, ==, NULL);
194 ASSERT3P(dn->dn_bonus, ==, NULL);
195 ASSERT(!dn->dn_have_spill);
196 ASSERT3P(dn->dn_zio, ==, NULL);
197 ASSERT0(dn->dn_oldused);
198 ASSERT0(dn->dn_oldflags);
199 ASSERT0(dn->dn_olduid);
200 ASSERT0(dn->dn_oldgid);
201 ASSERT0(dn->dn_newuid);
202 ASSERT0(dn->dn_newgid);
203 ASSERT0(dn->dn_id_flags);
204
205 ASSERT0(dn->dn_dbufs_count);
206 avl_destroy(&dn->dn_dbufs);
207 }
208
209 void
dnode_init(void)210 dnode_init(void)
211 {
212 ASSERT(dnode_cache == NULL);
213 dnode_cache = kmem_cache_create("dnode_t",
214 sizeof (dnode_t),
215 0, dnode_cons, dnode_dest, NULL, NULL, NULL, 0);
216 #ifdef _KERNEL
217 kmem_cache_set_move(dnode_cache, dnode_move);
218 #endif /* _KERNEL */
219 }
220
221 void
dnode_fini(void)222 dnode_fini(void)
223 {
224 kmem_cache_destroy(dnode_cache);
225 dnode_cache = NULL;
226 }
227
228
229 #ifdef ZFS_DEBUG
230 void
dnode_verify(dnode_t * dn)231 dnode_verify(dnode_t *dn)
232 {
233 int drop_struct_lock = FALSE;
234
235 ASSERT(dn->dn_phys);
236 ASSERT(dn->dn_objset);
237 ASSERT(dn->dn_handle->dnh_dnode == dn);
238
239 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
240
241 if (!(zfs_flags & ZFS_DEBUG_DNODE_VERIFY))
242 return;
243
244 if (!RW_WRITE_HELD(&dn->dn_struct_rwlock)) {
245 rw_enter(&dn->dn_struct_rwlock, RW_READER);
246 drop_struct_lock = TRUE;
247 }
248 if (dn->dn_phys->dn_type != DMU_OT_NONE || dn->dn_allocated_txg != 0) {
249 int i;
250 ASSERT3U(dn->dn_indblkshift, >=, 0);
251 ASSERT3U(dn->dn_indblkshift, <=, SPA_MAXBLOCKSHIFT);
252 if (dn->dn_datablkshift) {
253 ASSERT3U(dn->dn_datablkshift, >=, SPA_MINBLOCKSHIFT);
254 ASSERT3U(dn->dn_datablkshift, <=, SPA_MAXBLOCKSHIFT);
255 ASSERT3U(1<<dn->dn_datablkshift, ==, dn->dn_datablksz);
256 }
257 ASSERT3U(dn->dn_nlevels, <=, 30);
258 ASSERT(DMU_OT_IS_VALID(dn->dn_type));
259 ASSERT3U(dn->dn_nblkptr, >=, 1);
260 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
261 ASSERT3U(dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
262 ASSERT3U(dn->dn_datablksz, ==,
263 dn->dn_datablkszsec << SPA_MINBLOCKSHIFT);
264 ASSERT3U(ISP2(dn->dn_datablksz), ==, dn->dn_datablkshift != 0);
265 ASSERT3U((dn->dn_nblkptr - 1) * sizeof (blkptr_t) +
266 dn->dn_bonuslen, <=, DN_MAX_BONUSLEN);
267 for (i = 0; i < TXG_SIZE; i++) {
268 ASSERT3U(dn->dn_next_nlevels[i], <=, dn->dn_nlevels);
269 }
270 }
271 if (dn->dn_phys->dn_type != DMU_OT_NONE)
272 ASSERT3U(dn->dn_phys->dn_nlevels, <=, dn->dn_nlevels);
273 ASSERT(DMU_OBJECT_IS_SPECIAL(dn->dn_object) || dn->dn_dbuf != NULL);
274 if (dn->dn_dbuf != NULL) {
275 ASSERT3P(dn->dn_phys, ==,
276 (dnode_phys_t *)dn->dn_dbuf->db.db_data +
277 (dn->dn_object % (dn->dn_dbuf->db.db_size >> DNODE_SHIFT)));
278 }
279 if (drop_struct_lock)
280 rw_exit(&dn->dn_struct_rwlock);
281 }
282 #endif
283
284 void
dnode_byteswap(dnode_phys_t * dnp)285 dnode_byteswap(dnode_phys_t *dnp)
286 {
287 uint64_t *buf64 = (void*)&dnp->dn_blkptr;
288 int i;
289
290 if (dnp->dn_type == DMU_OT_NONE) {
291 bzero(dnp, sizeof (dnode_phys_t));
292 return;
293 }
294
295 dnp->dn_datablkszsec = BSWAP_16(dnp->dn_datablkszsec);
296 dnp->dn_bonuslen = BSWAP_16(dnp->dn_bonuslen);
297 dnp->dn_maxblkid = BSWAP_64(dnp->dn_maxblkid);
298 dnp->dn_used = BSWAP_64(dnp->dn_used);
299
300 /*
301 * dn_nblkptr is only one byte, so it's OK to read it in either
302 * byte order. We can't read dn_bouslen.
303 */
304 ASSERT(dnp->dn_indblkshift <= SPA_MAXBLOCKSHIFT);
305 ASSERT(dnp->dn_nblkptr <= DN_MAX_NBLKPTR);
306 for (i = 0; i < dnp->dn_nblkptr * sizeof (blkptr_t)/8; i++)
307 buf64[i] = BSWAP_64(buf64[i]);
308
309 /*
310 * OK to check dn_bonuslen for zero, because it won't matter if
311 * we have the wrong byte order. This is necessary because the
312 * dnode dnode is smaller than a regular dnode.
313 */
314 if (dnp->dn_bonuslen != 0) {
315 /*
316 * Note that the bonus length calculated here may be
317 * longer than the actual bonus buffer. This is because
318 * we always put the bonus buffer after the last block
319 * pointer (instead of packing it against the end of the
320 * dnode buffer).
321 */
322 int off = (dnp->dn_nblkptr-1) * sizeof (blkptr_t);
323 size_t len = DN_MAX_BONUSLEN - off;
324 ASSERT(DMU_OT_IS_VALID(dnp->dn_bonustype));
325 dmu_object_byteswap_t byteswap =
326 DMU_OT_BYTESWAP(dnp->dn_bonustype);
327 dmu_ot_byteswap[byteswap].ob_func(dnp->dn_bonus + off, len);
328 }
329
330 /* Swap SPILL block if we have one */
331 if (dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR)
332 byteswap_uint64_array(&dnp->dn_spill, sizeof (blkptr_t));
333
334 }
335
336 void
dnode_buf_byteswap(void * vbuf,size_t size)337 dnode_buf_byteswap(void *vbuf, size_t size)
338 {
339 dnode_phys_t *buf = vbuf;
340 int i;
341
342 ASSERT3U(sizeof (dnode_phys_t), ==, (1<<DNODE_SHIFT));
343 ASSERT((size & (sizeof (dnode_phys_t)-1)) == 0);
344
345 size >>= DNODE_SHIFT;
346 for (i = 0; i < size; i++) {
347 dnode_byteswap(buf);
348 buf++;
349 }
350 }
351
352 void
dnode_setbonuslen(dnode_t * dn,int newsize,dmu_tx_t * tx)353 dnode_setbonuslen(dnode_t *dn, int newsize, dmu_tx_t *tx)
354 {
355 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
356
357 dnode_setdirty(dn, tx);
358 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
359 ASSERT3U(newsize, <=, DN_MAX_BONUSLEN -
360 (dn->dn_nblkptr-1) * sizeof (blkptr_t));
361 dn->dn_bonuslen = newsize;
362 if (newsize == 0)
363 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = DN_ZERO_BONUSLEN;
364 else
365 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
366 rw_exit(&dn->dn_struct_rwlock);
367 }
368
369 void
dnode_setbonus_type(dnode_t * dn,dmu_object_type_t newtype,dmu_tx_t * tx)370 dnode_setbonus_type(dnode_t *dn, dmu_object_type_t newtype, dmu_tx_t *tx)
371 {
372 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
373 dnode_setdirty(dn, tx);
374 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
375 dn->dn_bonustype = newtype;
376 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
377 rw_exit(&dn->dn_struct_rwlock);
378 }
379
380 void
dnode_rm_spill(dnode_t * dn,dmu_tx_t * tx)381 dnode_rm_spill(dnode_t *dn, dmu_tx_t *tx)
382 {
383 ASSERT3U(refcount_count(&dn->dn_holds), >=, 1);
384 ASSERT(RW_WRITE_HELD(&dn->dn_struct_rwlock));
385 dnode_setdirty(dn, tx);
386 dn->dn_rm_spillblk[tx->tx_txg&TXG_MASK] = DN_KILL_SPILLBLK;
387 dn->dn_have_spill = B_FALSE;
388 }
389
390 static void
dnode_setdblksz(dnode_t * dn,int size)391 dnode_setdblksz(dnode_t *dn, int size)
392 {
393 ASSERT0(P2PHASE(size, SPA_MINBLOCKSIZE));
394 ASSERT3U(size, <=, SPA_MAXBLOCKSIZE);
395 ASSERT3U(size, >=, SPA_MINBLOCKSIZE);
396 ASSERT3U(size >> SPA_MINBLOCKSHIFT, <,
397 1<<(sizeof (dn->dn_phys->dn_datablkszsec) * 8));
398 dn->dn_datablksz = size;
399 dn->dn_datablkszsec = size >> SPA_MINBLOCKSHIFT;
400 dn->dn_datablkshift = ISP2(size) ? highbit64(size - 1) : 0;
401 }
402
403 static dnode_t *
dnode_create(objset_t * os,dnode_phys_t * dnp,dmu_buf_impl_t * db,uint64_t object,dnode_handle_t * dnh)404 dnode_create(objset_t *os, dnode_phys_t *dnp, dmu_buf_impl_t *db,
405 uint64_t object, dnode_handle_t *dnh)
406 {
407 dnode_t *dn;
408
409 dn = kmem_cache_alloc(dnode_cache, KM_SLEEP);
410 #ifdef _KERNEL
411 ASSERT(!POINTER_IS_VALID(dn->dn_objset));
412 #endif /* _KERNEL */
413 dn->dn_moved = 0;
414
415 /*
416 * Defer setting dn_objset until the dnode is ready to be a candidate
417 * for the dnode_move() callback.
418 */
419 dn->dn_object = object;
420 dn->dn_dbuf = db;
421 dn->dn_handle = dnh;
422 dn->dn_phys = dnp;
423
424 if (dnp->dn_datablkszsec) {
425 dnode_setdblksz(dn, dnp->dn_datablkszsec << SPA_MINBLOCKSHIFT);
426 } else {
427 dn->dn_datablksz = 0;
428 dn->dn_datablkszsec = 0;
429 dn->dn_datablkshift = 0;
430 }
431 dn->dn_indblkshift = dnp->dn_indblkshift;
432 dn->dn_nlevels = dnp->dn_nlevels;
433 dn->dn_type = dnp->dn_type;
434 dn->dn_nblkptr = dnp->dn_nblkptr;
435 dn->dn_checksum = dnp->dn_checksum;
436 dn->dn_compress = dnp->dn_compress;
437 dn->dn_bonustype = dnp->dn_bonustype;
438 dn->dn_bonuslen = dnp->dn_bonuslen;
439 dn->dn_maxblkid = dnp->dn_maxblkid;
440 dn->dn_have_spill = ((dnp->dn_flags & DNODE_FLAG_SPILL_BLKPTR) != 0);
441 dn->dn_id_flags = 0;
442
443 dmu_zfetch_init(&dn->dn_zfetch, dn);
444
445 ASSERT(DMU_OT_IS_VALID(dn->dn_phys->dn_type));
446
447 mutex_enter(&os->os_lock);
448 if (dnh->dnh_dnode != NULL) {
449 /* Lost the allocation race. */
450 mutex_exit(&os->os_lock);
451 kmem_cache_free(dnode_cache, dn);
452 return (dnh->dnh_dnode);
453 }
454
455 /*
456 * Exclude special dnodes from os_dnodes so an empty os_dnodes
457 * signifies that the special dnodes have no references from
458 * their children (the entries in os_dnodes). This allows
459 * dnode_destroy() to easily determine if the last child has
460 * been removed and then complete eviction of the objset.
461 */
462 if (!DMU_OBJECT_IS_SPECIAL(object))
463 list_insert_head(&os->os_dnodes, dn);
464 membar_producer();
465
466 /*
467 * Everything else must be valid before assigning dn_objset
468 * makes the dnode eligible for dnode_move().
469 */
470 dn->dn_objset = os;
471
472 dnh->dnh_dnode = dn;
473 mutex_exit(&os->os_lock);
474
475 arc_space_consume(sizeof (dnode_t), ARC_SPACE_OTHER);
476 return (dn);
477 }
478
479 /*
480 * Caller must be holding the dnode handle, which is released upon return.
481 */
482 static void
dnode_destroy(dnode_t * dn)483 dnode_destroy(dnode_t *dn)
484 {
485 objset_t *os = dn->dn_objset;
486 boolean_t complete_os_eviction = B_FALSE;
487
488 ASSERT((dn->dn_id_flags & DN_ID_NEW_EXIST) == 0);
489
490 mutex_enter(&os->os_lock);
491 POINTER_INVALIDATE(&dn->dn_objset);
492 if (!DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
493 list_remove(&os->os_dnodes, dn);
494 complete_os_eviction =
495 list_is_empty(&os->os_dnodes) &&
496 list_link_active(&os->os_evicting_node);
497 }
498 mutex_exit(&os->os_lock);
499
500 /* the dnode can no longer move, so we can release the handle */
501 zrl_remove(&dn->dn_handle->dnh_zrlock);
502
503 dn->dn_allocated_txg = 0;
504 dn->dn_free_txg = 0;
505 dn->dn_assigned_txg = 0;
506
507 dn->dn_dirtyctx = 0;
508 if (dn->dn_dirtyctx_firstset != NULL) {
509 kmem_free(dn->dn_dirtyctx_firstset, 1);
510 dn->dn_dirtyctx_firstset = NULL;
511 }
512 if (dn->dn_bonus != NULL) {
513 mutex_enter(&dn->dn_bonus->db_mtx);
514 dbuf_destroy(dn->dn_bonus);
515 dn->dn_bonus = NULL;
516 }
517 dn->dn_zio = NULL;
518
519 dn->dn_have_spill = B_FALSE;
520 dn->dn_oldused = 0;
521 dn->dn_oldflags = 0;
522 dn->dn_olduid = 0;
523 dn->dn_oldgid = 0;
524 dn->dn_newuid = 0;
525 dn->dn_newgid = 0;
526 dn->dn_id_flags = 0;
527
528 dmu_zfetch_fini(&dn->dn_zfetch);
529 kmem_cache_free(dnode_cache, dn);
530 arc_space_return(sizeof (dnode_t), ARC_SPACE_OTHER);
531
532 if (complete_os_eviction)
533 dmu_objset_evict_done(os);
534 }
535
536 void
dnode_allocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,int ibs,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)537 dnode_allocate(dnode_t *dn, dmu_object_type_t ot, int blocksize, int ibs,
538 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
539 {
540 int i;
541
542 ASSERT3U(blocksize, <=,
543 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
544 if (blocksize == 0)
545 blocksize = 1 << zfs_default_bs;
546 else
547 blocksize = P2ROUNDUP(blocksize, SPA_MINBLOCKSIZE);
548
549 if (ibs == 0)
550 ibs = zfs_default_ibs;
551
552 ibs = MIN(MAX(ibs, DN_MIN_INDBLKSHIFT), DN_MAX_INDBLKSHIFT);
553
554 dprintf("os=%p obj=%llu txg=%llu blocksize=%d ibs=%d\n", dn->dn_objset,
555 dn->dn_object, tx->tx_txg, blocksize, ibs);
556
557 ASSERT(dn->dn_type == DMU_OT_NONE);
558 ASSERT(bcmp(dn->dn_phys, &dnode_phys_zero, sizeof (dnode_phys_t)) == 0);
559 ASSERT(dn->dn_phys->dn_type == DMU_OT_NONE);
560 ASSERT(ot != DMU_OT_NONE);
561 ASSERT(DMU_OT_IS_VALID(ot));
562 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
563 (bonustype == DMU_OT_SA && bonuslen == 0) ||
564 (bonustype != DMU_OT_NONE && bonuslen != 0));
565 ASSERT(DMU_OT_IS_VALID(bonustype));
566 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
567 ASSERT(dn->dn_type == DMU_OT_NONE);
568 ASSERT0(dn->dn_maxblkid);
569 ASSERT0(dn->dn_allocated_txg);
570 ASSERT0(dn->dn_assigned_txg);
571 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
572 ASSERT3U(refcount_count(&dn->dn_holds), <=, 1);
573 ASSERT(avl_is_empty(&dn->dn_dbufs));
574
575 for (i = 0; i < TXG_SIZE; i++) {
576 ASSERT0(dn->dn_next_nblkptr[i]);
577 ASSERT0(dn->dn_next_nlevels[i]);
578 ASSERT0(dn->dn_next_indblkshift[i]);
579 ASSERT0(dn->dn_next_bonuslen[i]);
580 ASSERT0(dn->dn_next_bonustype[i]);
581 ASSERT0(dn->dn_rm_spillblk[i]);
582 ASSERT0(dn->dn_next_blksz[i]);
583 ASSERT(!list_link_active(&dn->dn_dirty_link[i]));
584 ASSERT3P(list_head(&dn->dn_dirty_records[i]), ==, NULL);
585 ASSERT3P(dn->dn_free_ranges[i], ==, NULL);
586 }
587
588 dn->dn_type = ot;
589 dnode_setdblksz(dn, blocksize);
590 dn->dn_indblkshift = ibs;
591 dn->dn_nlevels = 1;
592 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
593 dn->dn_nblkptr = 1;
594 else
595 dn->dn_nblkptr = 1 +
596 ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
597 dn->dn_bonustype = bonustype;
598 dn->dn_bonuslen = bonuslen;
599 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
600 dn->dn_compress = ZIO_COMPRESS_INHERIT;
601 dn->dn_dirtyctx = 0;
602
603 dn->dn_free_txg = 0;
604 if (dn->dn_dirtyctx_firstset) {
605 kmem_free(dn->dn_dirtyctx_firstset, 1);
606 dn->dn_dirtyctx_firstset = NULL;
607 }
608
609 dn->dn_allocated_txg = tx->tx_txg;
610 dn->dn_id_flags = 0;
611
612 dnode_setdirty(dn, tx);
613 dn->dn_next_indblkshift[tx->tx_txg & TXG_MASK] = ibs;
614 dn->dn_next_bonuslen[tx->tx_txg & TXG_MASK] = dn->dn_bonuslen;
615 dn->dn_next_bonustype[tx->tx_txg & TXG_MASK] = dn->dn_bonustype;
616 dn->dn_next_blksz[tx->tx_txg & TXG_MASK] = dn->dn_datablksz;
617 }
618
619 void
dnode_reallocate(dnode_t * dn,dmu_object_type_t ot,int blocksize,dmu_object_type_t bonustype,int bonuslen,dmu_tx_t * tx)620 dnode_reallocate(dnode_t *dn, dmu_object_type_t ot, int blocksize,
621 dmu_object_type_t bonustype, int bonuslen, dmu_tx_t *tx)
622 {
623 int nblkptr;
624
625 ASSERT3U(blocksize, >=, SPA_MINBLOCKSIZE);
626 ASSERT3U(blocksize, <=,
627 spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
628 ASSERT0(blocksize % SPA_MINBLOCKSIZE);
629 ASSERT(dn->dn_object != DMU_META_DNODE_OBJECT || dmu_tx_private_ok(tx));
630 ASSERT(tx->tx_txg != 0);
631 ASSERT((bonustype == DMU_OT_NONE && bonuslen == 0) ||
632 (bonustype != DMU_OT_NONE && bonuslen != 0) ||
633 (bonustype == DMU_OT_SA && bonuslen == 0));
634 ASSERT(DMU_OT_IS_VALID(bonustype));
635 ASSERT3U(bonuslen, <=, DN_MAX_BONUSLEN);
636
637 /* clean up any unreferenced dbufs */
638 dnode_evict_dbufs(dn);
639
640 dn->dn_id_flags = 0;
641
642 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
643 dnode_setdirty(dn, tx);
644 if (dn->dn_datablksz != blocksize) {
645 /* change blocksize */
646 ASSERT(dn->dn_maxblkid == 0 &&
647 (BP_IS_HOLE(&dn->dn_phys->dn_blkptr[0]) ||
648 dnode_block_freed(dn, 0)));
649 dnode_setdblksz(dn, blocksize);
650 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = blocksize;
651 }
652 if (dn->dn_bonuslen != bonuslen)
653 dn->dn_next_bonuslen[tx->tx_txg&TXG_MASK] = bonuslen;
654
655 if (bonustype == DMU_OT_SA) /* Maximize bonus space for SA */
656 nblkptr = 1;
657 else
658 nblkptr = 1 + ((DN_MAX_BONUSLEN - bonuslen) >> SPA_BLKPTRSHIFT);
659 if (dn->dn_bonustype != bonustype)
660 dn->dn_next_bonustype[tx->tx_txg&TXG_MASK] = bonustype;
661 if (dn->dn_nblkptr != nblkptr)
662 dn->dn_next_nblkptr[tx->tx_txg&TXG_MASK] = nblkptr;
663 if (dn->dn_phys->dn_flags & DNODE_FLAG_SPILL_BLKPTR) {
664 dbuf_rm_spill(dn, tx);
665 dnode_rm_spill(dn, tx);
666 }
667 rw_exit(&dn->dn_struct_rwlock);
668
669 /* change type */
670 dn->dn_type = ot;
671
672 /* change bonus size and type */
673 mutex_enter(&dn->dn_mtx);
674 dn->dn_bonustype = bonustype;
675 dn->dn_bonuslen = bonuslen;
676 dn->dn_nblkptr = nblkptr;
677 dn->dn_checksum = ZIO_CHECKSUM_INHERIT;
678 dn->dn_compress = ZIO_COMPRESS_INHERIT;
679 ASSERT3U(dn->dn_nblkptr, <=, DN_MAX_NBLKPTR);
680
681 /* fix up the bonus db_size */
682 if (dn->dn_bonus) {
683 dn->dn_bonus->db.db_size =
684 DN_MAX_BONUSLEN - (dn->dn_nblkptr-1) * sizeof (blkptr_t);
685 ASSERT(dn->dn_bonuslen <= dn->dn_bonus->db.db_size);
686 }
687
688 dn->dn_allocated_txg = tx->tx_txg;
689 mutex_exit(&dn->dn_mtx);
690 }
691
692 #ifdef DNODE_STATS
693 static struct {
694 uint64_t dms_dnode_invalid;
695 uint64_t dms_dnode_recheck1;
696 uint64_t dms_dnode_recheck2;
697 uint64_t dms_dnode_special;
698 uint64_t dms_dnode_handle;
699 uint64_t dms_dnode_rwlock;
700 uint64_t dms_dnode_active;
701 } dnode_move_stats;
702 #endif /* DNODE_STATS */
703
704 #ifdef _KERNEL
705 static void
dnode_move_impl(dnode_t * odn,dnode_t * ndn)706 dnode_move_impl(dnode_t *odn, dnode_t *ndn)
707 {
708 int i;
709
710 ASSERT(!RW_LOCK_HELD(&odn->dn_struct_rwlock));
711 ASSERT(MUTEX_NOT_HELD(&odn->dn_mtx));
712 ASSERT(MUTEX_NOT_HELD(&odn->dn_dbufs_mtx));
713 ASSERT(!RW_LOCK_HELD(&odn->dn_zfetch.zf_rwlock));
714
715 /* Copy fields. */
716 ndn->dn_objset = odn->dn_objset;
717 ndn->dn_object = odn->dn_object;
718 ndn->dn_dbuf = odn->dn_dbuf;
719 ndn->dn_handle = odn->dn_handle;
720 ndn->dn_phys = odn->dn_phys;
721 ndn->dn_type = odn->dn_type;
722 ndn->dn_bonuslen = odn->dn_bonuslen;
723 ndn->dn_bonustype = odn->dn_bonustype;
724 ndn->dn_nblkptr = odn->dn_nblkptr;
725 ndn->dn_checksum = odn->dn_checksum;
726 ndn->dn_compress = odn->dn_compress;
727 ndn->dn_nlevels = odn->dn_nlevels;
728 ndn->dn_indblkshift = odn->dn_indblkshift;
729 ndn->dn_datablkshift = odn->dn_datablkshift;
730 ndn->dn_datablkszsec = odn->dn_datablkszsec;
731 ndn->dn_datablksz = odn->dn_datablksz;
732 ndn->dn_maxblkid = odn->dn_maxblkid;
733 bcopy(&odn->dn_next_type[0], &ndn->dn_next_type[0],
734 sizeof (odn->dn_next_type));
735 bcopy(&odn->dn_next_nblkptr[0], &ndn->dn_next_nblkptr[0],
736 sizeof (odn->dn_next_nblkptr));
737 bcopy(&odn->dn_next_nlevels[0], &ndn->dn_next_nlevels[0],
738 sizeof (odn->dn_next_nlevels));
739 bcopy(&odn->dn_next_indblkshift[0], &ndn->dn_next_indblkshift[0],
740 sizeof (odn->dn_next_indblkshift));
741 bcopy(&odn->dn_next_bonustype[0], &ndn->dn_next_bonustype[0],
742 sizeof (odn->dn_next_bonustype));
743 bcopy(&odn->dn_rm_spillblk[0], &ndn->dn_rm_spillblk[0],
744 sizeof (odn->dn_rm_spillblk));
745 bcopy(&odn->dn_next_bonuslen[0], &ndn->dn_next_bonuslen[0],
746 sizeof (odn->dn_next_bonuslen));
747 bcopy(&odn->dn_next_blksz[0], &ndn->dn_next_blksz[0],
748 sizeof (odn->dn_next_blksz));
749 for (i = 0; i < TXG_SIZE; i++) {
750 list_move_tail(&ndn->dn_dirty_records[i],
751 &odn->dn_dirty_records[i]);
752 }
753 bcopy(&odn->dn_free_ranges[0], &ndn->dn_free_ranges[0],
754 sizeof (odn->dn_free_ranges));
755 ndn->dn_allocated_txg = odn->dn_allocated_txg;
756 ndn->dn_free_txg = odn->dn_free_txg;
757 ndn->dn_assigned_txg = odn->dn_assigned_txg;
758 ndn->dn_dirtyctx = odn->dn_dirtyctx;
759 ndn->dn_dirtyctx_firstset = odn->dn_dirtyctx_firstset;
760 ASSERT(refcount_count(&odn->dn_tx_holds) == 0);
761 refcount_transfer(&ndn->dn_holds, &odn->dn_holds);
762 ASSERT(avl_is_empty(&ndn->dn_dbufs));
763 avl_swap(&ndn->dn_dbufs, &odn->dn_dbufs);
764 ndn->dn_dbufs_count = odn->dn_dbufs_count;
765 ndn->dn_bonus = odn->dn_bonus;
766 ndn->dn_have_spill = odn->dn_have_spill;
767 ndn->dn_zio = odn->dn_zio;
768 ndn->dn_oldused = odn->dn_oldused;
769 ndn->dn_oldflags = odn->dn_oldflags;
770 ndn->dn_olduid = odn->dn_olduid;
771 ndn->dn_oldgid = odn->dn_oldgid;
772 ndn->dn_newuid = odn->dn_newuid;
773 ndn->dn_newgid = odn->dn_newgid;
774 ndn->dn_id_flags = odn->dn_id_flags;
775 dmu_zfetch_init(&ndn->dn_zfetch, NULL);
776 list_move_tail(&ndn->dn_zfetch.zf_stream, &odn->dn_zfetch.zf_stream);
777 ndn->dn_zfetch.zf_dnode = odn->dn_zfetch.zf_dnode;
778
779 /*
780 * Update back pointers. Updating the handle fixes the back pointer of
781 * every descendant dbuf as well as the bonus dbuf.
782 */
783 ASSERT(ndn->dn_handle->dnh_dnode == odn);
784 ndn->dn_handle->dnh_dnode = ndn;
785 if (ndn->dn_zfetch.zf_dnode == odn) {
786 ndn->dn_zfetch.zf_dnode = ndn;
787 }
788
789 /*
790 * Invalidate the original dnode by clearing all of its back pointers.
791 */
792 odn->dn_dbuf = NULL;
793 odn->dn_handle = NULL;
794 avl_create(&odn->dn_dbufs, dbuf_compare, sizeof (dmu_buf_impl_t),
795 offsetof(dmu_buf_impl_t, db_link));
796 odn->dn_dbufs_count = 0;
797 odn->dn_bonus = NULL;
798 odn->dn_zfetch.zf_dnode = NULL;
799
800 /*
801 * Set the low bit of the objset pointer to ensure that dnode_move()
802 * recognizes the dnode as invalid in any subsequent callback.
803 */
804 POINTER_INVALIDATE(&odn->dn_objset);
805
806 /*
807 * Satisfy the destructor.
808 */
809 for (i = 0; i < TXG_SIZE; i++) {
810 list_create(&odn->dn_dirty_records[i],
811 sizeof (dbuf_dirty_record_t),
812 offsetof(dbuf_dirty_record_t, dr_dirty_node));
813 odn->dn_free_ranges[i] = NULL;
814 odn->dn_next_nlevels[i] = 0;
815 odn->dn_next_indblkshift[i] = 0;
816 odn->dn_next_bonustype[i] = 0;
817 odn->dn_rm_spillblk[i] = 0;
818 odn->dn_next_bonuslen[i] = 0;
819 odn->dn_next_blksz[i] = 0;
820 }
821 odn->dn_allocated_txg = 0;
822 odn->dn_free_txg = 0;
823 odn->dn_assigned_txg = 0;
824 odn->dn_dirtyctx = 0;
825 odn->dn_dirtyctx_firstset = NULL;
826 odn->dn_have_spill = B_FALSE;
827 odn->dn_zio = NULL;
828 odn->dn_oldused = 0;
829 odn->dn_oldflags = 0;
830 odn->dn_olduid = 0;
831 odn->dn_oldgid = 0;
832 odn->dn_newuid = 0;
833 odn->dn_newgid = 0;
834 odn->dn_id_flags = 0;
835
836 /*
837 * Mark the dnode.
838 */
839 ndn->dn_moved = 1;
840 odn->dn_moved = (uint8_t)-1;
841 }
842
843 #ifdef illumos
844 /*ARGSUSED*/
845 static kmem_cbrc_t
dnode_move(void * buf,void * newbuf,size_t size,void * arg)846 dnode_move(void *buf, void *newbuf, size_t size, void *arg)
847 {
848 dnode_t *odn = buf, *ndn = newbuf;
849 objset_t *os;
850 int64_t refcount;
851 uint32_t dbufs;
852
853 /*
854 * The dnode is on the objset's list of known dnodes if the objset
855 * pointer is valid. We set the low bit of the objset pointer when
856 * freeing the dnode to invalidate it, and the memory patterns written
857 * by kmem (baddcafe and deadbeef) set at least one of the two low bits.
858 * A newly created dnode sets the objset pointer last of all to indicate
859 * that the dnode is known and in a valid state to be moved by this
860 * function.
861 */
862 os = odn->dn_objset;
863 if (!POINTER_IS_VALID(os)) {
864 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_invalid);
865 return (KMEM_CBRC_DONT_KNOW);
866 }
867
868 /*
869 * Ensure that the objset does not go away during the move.
870 */
871 rw_enter(&os_lock, RW_WRITER);
872 if (os != odn->dn_objset) {
873 rw_exit(&os_lock);
874 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck1);
875 return (KMEM_CBRC_DONT_KNOW);
876 }
877
878 /*
879 * If the dnode is still valid, then so is the objset. We know that no
880 * valid objset can be freed while we hold os_lock, so we can safely
881 * ensure that the objset remains in use.
882 */
883 mutex_enter(&os->os_lock);
884
885 /*
886 * Recheck the objset pointer in case the dnode was removed just before
887 * acquiring the lock.
888 */
889 if (os != odn->dn_objset) {
890 mutex_exit(&os->os_lock);
891 rw_exit(&os_lock);
892 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_recheck2);
893 return (KMEM_CBRC_DONT_KNOW);
894 }
895
896 /*
897 * At this point we know that as long as we hold os->os_lock, the dnode
898 * cannot be freed and fields within the dnode can be safely accessed.
899 * The objset listing this dnode cannot go away as long as this dnode is
900 * on its list.
901 */
902 rw_exit(&os_lock);
903 if (DMU_OBJECT_IS_SPECIAL(odn->dn_object)) {
904 mutex_exit(&os->os_lock);
905 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_special);
906 return (KMEM_CBRC_NO);
907 }
908 ASSERT(odn->dn_dbuf != NULL); /* only "special" dnodes have no parent */
909
910 /*
911 * Lock the dnode handle to prevent the dnode from obtaining any new
912 * holds. This also prevents the descendant dbufs and the bonus dbuf
913 * from accessing the dnode, so that we can discount their holds. The
914 * handle is safe to access because we know that while the dnode cannot
915 * go away, neither can its handle. Once we hold dnh_zrlock, we can
916 * safely move any dnode referenced only by dbufs.
917 */
918 if (!zrl_tryenter(&odn->dn_handle->dnh_zrlock)) {
919 mutex_exit(&os->os_lock);
920 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_handle);
921 return (KMEM_CBRC_LATER);
922 }
923
924 /*
925 * Ensure a consistent view of the dnode's holds and the dnode's dbufs.
926 * We need to guarantee that there is a hold for every dbuf in order to
927 * determine whether the dnode is actively referenced. Falsely matching
928 * a dbuf to an active hold would lead to an unsafe move. It's possible
929 * that a thread already having an active dnode hold is about to add a
930 * dbuf, and we can't compare hold and dbuf counts while the add is in
931 * progress.
932 */
933 if (!rw_tryenter(&odn->dn_struct_rwlock, RW_WRITER)) {
934 zrl_exit(&odn->dn_handle->dnh_zrlock);
935 mutex_exit(&os->os_lock);
936 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_rwlock);
937 return (KMEM_CBRC_LATER);
938 }
939
940 /*
941 * A dbuf may be removed (evicted) without an active dnode hold. In that
942 * case, the dbuf count is decremented under the handle lock before the
943 * dbuf's hold is released. This order ensures that if we count the hold
944 * after the dbuf is removed but before its hold is released, we will
945 * treat the unmatched hold as active and exit safely. If we count the
946 * hold before the dbuf is removed, the hold is discounted, and the
947 * removal is blocked until the move completes.
948 */
949 refcount = refcount_count(&odn->dn_holds);
950 ASSERT(refcount >= 0);
951 dbufs = odn->dn_dbufs_count;
952
953 /* We can't have more dbufs than dnode holds. */
954 ASSERT3U(dbufs, <=, refcount);
955 DTRACE_PROBE3(dnode__move, dnode_t *, odn, int64_t, refcount,
956 uint32_t, dbufs);
957
958 if (refcount > dbufs) {
959 rw_exit(&odn->dn_struct_rwlock);
960 zrl_exit(&odn->dn_handle->dnh_zrlock);
961 mutex_exit(&os->os_lock);
962 DNODE_STAT_ADD(dnode_move_stats.dms_dnode_active);
963 return (KMEM_CBRC_LATER);
964 }
965
966 rw_exit(&odn->dn_struct_rwlock);
967
968 /*
969 * At this point we know that anyone with a hold on the dnode is not
970 * actively referencing it. The dnode is known and in a valid state to
971 * move. We're holding the locks needed to execute the critical section.
972 */
973 dnode_move_impl(odn, ndn);
974
975 list_link_replace(&odn->dn_link, &ndn->dn_link);
976 /* If the dnode was safe to move, the refcount cannot have changed. */
977 ASSERT(refcount == refcount_count(&ndn->dn_holds));
978 ASSERT(dbufs == ndn->dn_dbufs_count);
979 zrl_exit(&ndn->dn_handle->dnh_zrlock); /* handle has moved */
980 mutex_exit(&os->os_lock);
981
982 return (KMEM_CBRC_YES);
983 }
984 #endif /* illumos */
985 #endif /* _KERNEL */
986
987 void
dnode_special_close(dnode_handle_t * dnh)988 dnode_special_close(dnode_handle_t *dnh)
989 {
990 dnode_t *dn = dnh->dnh_dnode;
991
992 /*
993 * Wait for final references to the dnode to clear. This can
994 * only happen if the arc is asyncronously evicting state that
995 * has a hold on this dnode while we are trying to evict this
996 * dnode.
997 */
998 while (refcount_count(&dn->dn_holds) > 0)
999 delay(1);
1000 ASSERT(dn->dn_dbuf == NULL ||
1001 dmu_buf_get_user(&dn->dn_dbuf->db) == NULL);
1002 zrl_add(&dnh->dnh_zrlock);
1003 dnode_destroy(dn); /* implicit zrl_remove() */
1004 zrl_destroy(&dnh->dnh_zrlock);
1005 dnh->dnh_dnode = NULL;
1006 }
1007
1008 void
dnode_special_open(objset_t * os,dnode_phys_t * dnp,uint64_t object,dnode_handle_t * dnh)1009 dnode_special_open(objset_t *os, dnode_phys_t *dnp, uint64_t object,
1010 dnode_handle_t *dnh)
1011 {
1012 dnode_t *dn;
1013
1014 dn = dnode_create(os, dnp, NULL, object, dnh);
1015 zrl_init(&dnh->dnh_zrlock);
1016 DNODE_VERIFY(dn);
1017 }
1018
1019 static void
dnode_buf_evict_async(void * dbu)1020 dnode_buf_evict_async(void *dbu)
1021 {
1022 dnode_children_t *children_dnodes = dbu;
1023 int i;
1024
1025 for (i = 0; i < children_dnodes->dnc_count; i++) {
1026 dnode_handle_t *dnh = &children_dnodes->dnc_children[i];
1027 dnode_t *dn;
1028
1029 /*
1030 * The dnode handle lock guards against the dnode moving to
1031 * another valid address, so there is no need here to guard
1032 * against changes to or from NULL.
1033 */
1034 if (dnh->dnh_dnode == NULL) {
1035 zrl_destroy(&dnh->dnh_zrlock);
1036 continue;
1037 }
1038
1039 zrl_add(&dnh->dnh_zrlock);
1040 dn = dnh->dnh_dnode;
1041 /*
1042 * If there are holds on this dnode, then there should
1043 * be holds on the dnode's containing dbuf as well; thus
1044 * it wouldn't be eligible for eviction and this function
1045 * would not have been called.
1046 */
1047 ASSERT(refcount_is_zero(&dn->dn_holds));
1048 ASSERT(refcount_is_zero(&dn->dn_tx_holds));
1049
1050 dnode_destroy(dn); /* implicit zrl_remove() */
1051 zrl_destroy(&dnh->dnh_zrlock);
1052 dnh->dnh_dnode = NULL;
1053 }
1054 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1055 children_dnodes->dnc_count * sizeof (dnode_handle_t));
1056 }
1057
1058 /*
1059 * errors:
1060 * EINVAL - invalid object number.
1061 * EIO - i/o error.
1062 * succeeds even for free dnodes.
1063 */
1064 int
dnode_hold_impl(objset_t * os,uint64_t object,int flag,void * tag,dnode_t ** dnp)1065 dnode_hold_impl(objset_t *os, uint64_t object, int flag,
1066 void *tag, dnode_t **dnp)
1067 {
1068 int epb, idx, err;
1069 int drop_struct_lock = FALSE;
1070 int type;
1071 uint64_t blk;
1072 dnode_t *mdn, *dn;
1073 dmu_buf_impl_t *db;
1074 dnode_children_t *children_dnodes;
1075 dnode_handle_t *dnh;
1076
1077 /*
1078 * If you are holding the spa config lock as writer, you shouldn't
1079 * be asking the DMU to do *anything* unless it's the root pool
1080 * which may require us to read from the root filesystem while
1081 * holding some (not all) of the locks as writer.
1082 */
1083 ASSERT(spa_config_held(os->os_spa, SCL_ALL, RW_WRITER) == 0 ||
1084 (spa_is_root(os->os_spa) &&
1085 spa_config_held(os->os_spa, SCL_STATE, RW_WRITER)));
1086
1087 ASSERT((flag & DNODE_MUST_BE_ALLOCATED) || (flag & DNODE_MUST_BE_FREE));
1088
1089 if (object == DMU_USERUSED_OBJECT || object == DMU_GROUPUSED_OBJECT) {
1090 dn = (object == DMU_USERUSED_OBJECT) ?
1091 DMU_USERUSED_DNODE(os) : DMU_GROUPUSED_DNODE(os);
1092 if (dn == NULL)
1093 return (SET_ERROR(ENOENT));
1094 type = dn->dn_type;
1095 if ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE)
1096 return (SET_ERROR(ENOENT));
1097 if ((flag & DNODE_MUST_BE_FREE) && type != DMU_OT_NONE)
1098 return (SET_ERROR(EEXIST));
1099 DNODE_VERIFY(dn);
1100 (void) refcount_add(&dn->dn_holds, tag);
1101 *dnp = dn;
1102 return (0);
1103 }
1104
1105 if (object == 0 || object >= DN_MAX_OBJECT)
1106 return (SET_ERROR(EINVAL));
1107
1108 mdn = DMU_META_DNODE(os);
1109 ASSERT(mdn->dn_object == DMU_META_DNODE_OBJECT);
1110
1111 DNODE_VERIFY(mdn);
1112
1113 if (!RW_WRITE_HELD(&mdn->dn_struct_rwlock)) {
1114 rw_enter(&mdn->dn_struct_rwlock, RW_READER);
1115 drop_struct_lock = TRUE;
1116 }
1117
1118 blk = dbuf_whichblock(mdn, 0, object * sizeof (dnode_phys_t));
1119
1120 db = dbuf_hold(mdn, blk, FTAG);
1121 if (drop_struct_lock)
1122 rw_exit(&mdn->dn_struct_rwlock);
1123 if (db == NULL)
1124 return (SET_ERROR(EIO));
1125 err = dbuf_read(db, NULL, DB_RF_CANFAIL);
1126 if (err) {
1127 dbuf_rele(db, FTAG);
1128 return (err);
1129 }
1130
1131 ASSERT3U(db->db.db_size, >=, 1<<DNODE_SHIFT);
1132 epb = db->db.db_size >> DNODE_SHIFT;
1133
1134 idx = object & (epb-1);
1135
1136 ASSERT(DB_DNODE(db)->dn_type == DMU_OT_DNODE);
1137 children_dnodes = dmu_buf_get_user(&db->db);
1138 if (children_dnodes == NULL) {
1139 int i;
1140 dnode_children_t *winner;
1141 children_dnodes = kmem_zalloc(sizeof (dnode_children_t) +
1142 epb * sizeof (dnode_handle_t), KM_SLEEP);
1143 children_dnodes->dnc_count = epb;
1144 dnh = &children_dnodes->dnc_children[0];
1145 for (i = 0; i < epb; i++) {
1146 zrl_init(&dnh[i].dnh_zrlock);
1147 }
1148 dmu_buf_init_user(&children_dnodes->dnc_dbu, NULL,
1149 dnode_buf_evict_async, NULL);
1150 winner = dmu_buf_set_user(&db->db, &children_dnodes->dnc_dbu);
1151 if (winner != NULL) {
1152
1153 for (i = 0; i < epb; i++) {
1154 zrl_destroy(&dnh[i].dnh_zrlock);
1155 }
1156
1157 kmem_free(children_dnodes, sizeof (dnode_children_t) +
1158 epb * sizeof (dnode_handle_t));
1159 children_dnodes = winner;
1160 }
1161 }
1162 ASSERT(children_dnodes->dnc_count == epb);
1163
1164 dnh = &children_dnodes->dnc_children[idx];
1165 zrl_add(&dnh->dnh_zrlock);
1166 dn = dnh->dnh_dnode;
1167 if (dn == NULL) {
1168 dnode_phys_t *phys = (dnode_phys_t *)db->db.db_data+idx;
1169
1170 dn = dnode_create(os, phys, db, object, dnh);
1171 }
1172
1173 mutex_enter(&dn->dn_mtx);
1174 type = dn->dn_type;
1175 if (dn->dn_free_txg ||
1176 ((flag & DNODE_MUST_BE_ALLOCATED) && type == DMU_OT_NONE) ||
1177 ((flag & DNODE_MUST_BE_FREE) &&
1178 (type != DMU_OT_NONE || !refcount_is_zero(&dn->dn_holds)))) {
1179 mutex_exit(&dn->dn_mtx);
1180 zrl_remove(&dnh->dnh_zrlock);
1181 dbuf_rele(db, FTAG);
1182 return ((flag & DNODE_MUST_BE_ALLOCATED) ? ENOENT : EEXIST);
1183 }
1184 if (refcount_add(&dn->dn_holds, tag) == 1)
1185 dbuf_add_ref(db, dnh);
1186 mutex_exit(&dn->dn_mtx);
1187
1188 /* Now we can rely on the hold to prevent the dnode from moving. */
1189 zrl_remove(&dnh->dnh_zrlock);
1190
1191 DNODE_VERIFY(dn);
1192 ASSERT3P(dn->dn_dbuf, ==, db);
1193 ASSERT3U(dn->dn_object, ==, object);
1194 dbuf_rele(db, FTAG);
1195
1196 *dnp = dn;
1197 return (0);
1198 }
1199
1200 /*
1201 * Return held dnode if the object is allocated, NULL if not.
1202 */
1203 int
dnode_hold(objset_t * os,uint64_t object,void * tag,dnode_t ** dnp)1204 dnode_hold(objset_t *os, uint64_t object, void *tag, dnode_t **dnp)
1205 {
1206 return (dnode_hold_impl(os, object, DNODE_MUST_BE_ALLOCATED, tag, dnp));
1207 }
1208
1209 /*
1210 * Can only add a reference if there is already at least one
1211 * reference on the dnode. Returns FALSE if unable to add a
1212 * new reference.
1213 */
1214 boolean_t
dnode_add_ref(dnode_t * dn,void * tag)1215 dnode_add_ref(dnode_t *dn, void *tag)
1216 {
1217 mutex_enter(&dn->dn_mtx);
1218 if (refcount_is_zero(&dn->dn_holds)) {
1219 mutex_exit(&dn->dn_mtx);
1220 return (FALSE);
1221 }
1222 VERIFY(1 < refcount_add(&dn->dn_holds, tag));
1223 mutex_exit(&dn->dn_mtx);
1224 return (TRUE);
1225 }
1226
1227 void
dnode_rele(dnode_t * dn,void * tag)1228 dnode_rele(dnode_t *dn, void *tag)
1229 {
1230 mutex_enter(&dn->dn_mtx);
1231 dnode_rele_and_unlock(dn, tag, B_FALSE);
1232 }
1233
1234 void
dnode_rele_and_unlock(dnode_t * dn,void * tag,boolean_t evicting)1235 dnode_rele_and_unlock(dnode_t *dn, void *tag, boolean_t evicting)
1236 {
1237 uint64_t refs;
1238 /* Get while the hold prevents the dnode from moving. */
1239 dmu_buf_impl_t *db = dn->dn_dbuf;
1240 dnode_handle_t *dnh = dn->dn_handle;
1241
1242 refs = refcount_remove(&dn->dn_holds, tag);
1243 mutex_exit(&dn->dn_mtx);
1244
1245 /*
1246 * It's unsafe to release the last hold on a dnode by dnode_rele() or
1247 * indirectly by dbuf_rele() while relying on the dnode handle to
1248 * prevent the dnode from moving, since releasing the last hold could
1249 * result in the dnode's parent dbuf evicting its dnode handles. For
1250 * that reason anyone calling dnode_rele() or dbuf_rele() without some
1251 * other direct or indirect hold on the dnode must first drop the dnode
1252 * handle.
1253 */
1254 ASSERT(refs > 0 || dnh->dnh_zrlock.zr_owner != curthread);
1255
1256 /* NOTE: the DNODE_DNODE does not have a dn_dbuf */
1257 if (refs == 0 && db != NULL) {
1258 /*
1259 * Another thread could add a hold to the dnode handle in
1260 * dnode_hold_impl() while holding the parent dbuf. Since the
1261 * hold on the parent dbuf prevents the handle from being
1262 * destroyed, the hold on the handle is OK. We can't yet assert
1263 * that the handle has zero references, but that will be
1264 * asserted anyway when the handle gets destroyed.
1265 */
1266 mutex_enter(&db->db_mtx);
1267 dbuf_rele_and_unlock(db, dnh, evicting);
1268 }
1269 }
1270
1271 void
dnode_setdirty(dnode_t * dn,dmu_tx_t * tx)1272 dnode_setdirty(dnode_t *dn, dmu_tx_t *tx)
1273 {
1274 objset_t *os = dn->dn_objset;
1275 uint64_t txg = tx->tx_txg;
1276
1277 if (DMU_OBJECT_IS_SPECIAL(dn->dn_object)) {
1278 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1279 return;
1280 }
1281
1282 DNODE_VERIFY(dn);
1283
1284 #ifdef ZFS_DEBUG
1285 mutex_enter(&dn->dn_mtx);
1286 ASSERT(dn->dn_phys->dn_type || dn->dn_allocated_txg);
1287 ASSERT(dn->dn_free_txg == 0 || dn->dn_free_txg >= txg);
1288 mutex_exit(&dn->dn_mtx);
1289 #endif
1290
1291 /*
1292 * Determine old uid/gid when necessary
1293 */
1294 dmu_objset_userquota_get_ids(dn, B_TRUE, tx);
1295
1296 multilist_t *dirtylist = os->os_dirty_dnodes[txg & TXG_MASK];
1297 multilist_sublist_t *mls = multilist_sublist_lock_obj(dirtylist, dn);
1298
1299 /*
1300 * If we are already marked dirty, we're done.
1301 */
1302 if (list_link_active(&dn->dn_dirty_link[txg & TXG_MASK])) {
1303 multilist_sublist_unlock(mls);
1304 return;
1305 }
1306
1307 ASSERT(!refcount_is_zero(&dn->dn_holds) ||
1308 !avl_is_empty(&dn->dn_dbufs));
1309 ASSERT(dn->dn_datablksz != 0);
1310 ASSERT0(dn->dn_next_bonuslen[txg&TXG_MASK]);
1311 ASSERT0(dn->dn_next_blksz[txg&TXG_MASK]);
1312 ASSERT0(dn->dn_next_bonustype[txg&TXG_MASK]);
1313
1314 dprintf_ds(os->os_dsl_dataset, "obj=%llu txg=%llu\n",
1315 dn->dn_object, txg);
1316
1317 multilist_sublist_insert_head(mls, dn);
1318
1319 multilist_sublist_unlock(mls);
1320
1321 /*
1322 * The dnode maintains a hold on its containing dbuf as
1323 * long as there are holds on it. Each instantiated child
1324 * dbuf maintains a hold on the dnode. When the last child
1325 * drops its hold, the dnode will drop its hold on the
1326 * containing dbuf. We add a "dirty hold" here so that the
1327 * dnode will hang around after we finish processing its
1328 * children.
1329 */
1330 VERIFY(dnode_add_ref(dn, (void *)(uintptr_t)tx->tx_txg));
1331
1332 (void) dbuf_dirty(dn->dn_dbuf, tx);
1333
1334 dsl_dataset_dirty(os->os_dsl_dataset, tx);
1335 }
1336
1337 void
dnode_free(dnode_t * dn,dmu_tx_t * tx)1338 dnode_free(dnode_t *dn, dmu_tx_t *tx)
1339 {
1340 mutex_enter(&dn->dn_mtx);
1341 if (dn->dn_type == DMU_OT_NONE || dn->dn_free_txg) {
1342 mutex_exit(&dn->dn_mtx);
1343 return;
1344 }
1345 dn->dn_free_txg = tx->tx_txg;
1346 mutex_exit(&dn->dn_mtx);
1347
1348 dnode_setdirty(dn, tx);
1349 }
1350
1351 /*
1352 * Try to change the block size for the indicated dnode. This can only
1353 * succeed if there are no blocks allocated or dirty beyond first block
1354 */
1355 int
dnode_set_blksz(dnode_t * dn,uint64_t size,int ibs,dmu_tx_t * tx)1356 dnode_set_blksz(dnode_t *dn, uint64_t size, int ibs, dmu_tx_t *tx)
1357 {
1358 dmu_buf_impl_t *db;
1359 int err;
1360
1361 ASSERT3U(size, <=, spa_maxblocksize(dmu_objset_spa(dn->dn_objset)));
1362 if (size == 0)
1363 size = SPA_MINBLOCKSIZE;
1364 else
1365 size = P2ROUNDUP(size, SPA_MINBLOCKSIZE);
1366
1367 if (ibs == dn->dn_indblkshift)
1368 ibs = 0;
1369
1370 if (size >> SPA_MINBLOCKSHIFT == dn->dn_datablkszsec && ibs == 0)
1371 return (0);
1372
1373 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1374
1375 /* Check for any allocated blocks beyond the first */
1376 if (dn->dn_maxblkid != 0)
1377 goto fail;
1378
1379 mutex_enter(&dn->dn_dbufs_mtx);
1380 for (db = avl_first(&dn->dn_dbufs); db != NULL;
1381 db = AVL_NEXT(&dn->dn_dbufs, db)) {
1382 if (db->db_blkid != 0 && db->db_blkid != DMU_BONUS_BLKID &&
1383 db->db_blkid != DMU_SPILL_BLKID) {
1384 mutex_exit(&dn->dn_dbufs_mtx);
1385 goto fail;
1386 }
1387 }
1388 mutex_exit(&dn->dn_dbufs_mtx);
1389
1390 if (ibs && dn->dn_nlevels != 1)
1391 goto fail;
1392
1393 /* resize the old block */
1394 err = dbuf_hold_impl(dn, 0, 0, TRUE, FALSE, FTAG, &db);
1395 if (err == 0)
1396 dbuf_new_size(db, size, tx);
1397 else if (err != ENOENT)
1398 goto fail;
1399
1400 dnode_setdblksz(dn, size);
1401 dnode_setdirty(dn, tx);
1402 dn->dn_next_blksz[tx->tx_txg&TXG_MASK] = size;
1403 if (ibs) {
1404 dn->dn_indblkshift = ibs;
1405 dn->dn_next_indblkshift[tx->tx_txg&TXG_MASK] = ibs;
1406 }
1407 /* rele after we have fixed the blocksize in the dnode */
1408 if (db)
1409 dbuf_rele(db, FTAG);
1410
1411 rw_exit(&dn->dn_struct_rwlock);
1412 return (0);
1413
1414 fail:
1415 rw_exit(&dn->dn_struct_rwlock);
1416 return (SET_ERROR(ENOTSUP));
1417 }
1418
1419 /* read-holding callers must not rely on the lock being continuously held */
1420 void
dnode_new_blkid(dnode_t * dn,uint64_t blkid,dmu_tx_t * tx,boolean_t have_read)1421 dnode_new_blkid(dnode_t *dn, uint64_t blkid, dmu_tx_t *tx, boolean_t have_read)
1422 {
1423 uint64_t txgoff = tx->tx_txg & TXG_MASK;
1424 int epbs, new_nlevels;
1425 uint64_t sz;
1426
1427 ASSERT(blkid != DMU_BONUS_BLKID);
1428
1429 ASSERT(have_read ?
1430 RW_READ_HELD(&dn->dn_struct_rwlock) :
1431 RW_WRITE_HELD(&dn->dn_struct_rwlock));
1432
1433 /*
1434 * if we have a read-lock, check to see if we need to do any work
1435 * before upgrading to a write-lock.
1436 */
1437 if (have_read) {
1438 if (blkid <= dn->dn_maxblkid)
1439 return;
1440
1441 if (!rw_tryupgrade(&dn->dn_struct_rwlock)) {
1442 rw_exit(&dn->dn_struct_rwlock);
1443 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1444 }
1445 }
1446
1447 if (blkid <= dn->dn_maxblkid)
1448 goto out;
1449
1450 dn->dn_maxblkid = blkid;
1451
1452 /*
1453 * Compute the number of levels necessary to support the new maxblkid.
1454 */
1455 new_nlevels = 1;
1456 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1457 for (sz = dn->dn_nblkptr;
1458 sz <= blkid && sz >= dn->dn_nblkptr; sz <<= epbs)
1459 new_nlevels++;
1460
1461 if (new_nlevels > dn->dn_nlevels) {
1462 int old_nlevels = dn->dn_nlevels;
1463 dmu_buf_impl_t *db;
1464 list_t *list;
1465 dbuf_dirty_record_t *new, *dr, *dr_next;
1466
1467 dn->dn_nlevels = new_nlevels;
1468
1469 ASSERT3U(new_nlevels, >, dn->dn_next_nlevels[txgoff]);
1470 dn->dn_next_nlevels[txgoff] = new_nlevels;
1471
1472 /* dirty the left indirects */
1473 db = dbuf_hold_level(dn, old_nlevels, 0, FTAG);
1474 ASSERT(db != NULL);
1475 new = dbuf_dirty(db, tx);
1476 dbuf_rele(db, FTAG);
1477
1478 /* transfer the dirty records to the new indirect */
1479 mutex_enter(&dn->dn_mtx);
1480 mutex_enter(&new->dt.di.dr_mtx);
1481 list = &dn->dn_dirty_records[txgoff];
1482 for (dr = list_head(list); dr; dr = dr_next) {
1483 dr_next = list_next(&dn->dn_dirty_records[txgoff], dr);
1484 if (dr->dr_dbuf->db_level != new_nlevels-1 &&
1485 dr->dr_dbuf->db_blkid != DMU_BONUS_BLKID &&
1486 dr->dr_dbuf->db_blkid != DMU_SPILL_BLKID) {
1487 ASSERT(dr->dr_dbuf->db_level == old_nlevels-1);
1488 list_remove(&dn->dn_dirty_records[txgoff], dr);
1489 list_insert_tail(&new->dt.di.dr_children, dr);
1490 dr->dr_parent = new;
1491 }
1492 }
1493 mutex_exit(&new->dt.di.dr_mtx);
1494 mutex_exit(&dn->dn_mtx);
1495 }
1496
1497 out:
1498 if (have_read)
1499 rw_downgrade(&dn->dn_struct_rwlock);
1500 }
1501
1502 static void
dnode_dirty_l1(dnode_t * dn,uint64_t l1blkid,dmu_tx_t * tx)1503 dnode_dirty_l1(dnode_t *dn, uint64_t l1blkid, dmu_tx_t *tx)
1504 {
1505 dmu_buf_impl_t *db = dbuf_hold_level(dn, 1, l1blkid, FTAG);
1506 if (db != NULL) {
1507 dmu_buf_will_dirty(&db->db, tx);
1508 dbuf_rele(db, FTAG);
1509 }
1510 }
1511
1512 /*
1513 * Dirty all the in-core level-1 dbufs in the range specified by start_blkid
1514 * and end_blkid.
1515 */
1516 static void
dnode_dirty_l1range(dnode_t * dn,uint64_t start_blkid,uint64_t end_blkid,dmu_tx_t * tx)1517 dnode_dirty_l1range(dnode_t *dn, uint64_t start_blkid, uint64_t end_blkid,
1518 dmu_tx_t *tx)
1519 {
1520 dmu_buf_impl_t db_search;
1521 dmu_buf_impl_t *db;
1522 avl_index_t where;
1523
1524 mutex_enter(&dn->dn_dbufs_mtx);
1525
1526 db_search.db_level = 1;
1527 db_search.db_blkid = start_blkid + 1;
1528 db_search.db_state = DB_SEARCH;
1529 for (;;) {
1530
1531 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1532 if (db == NULL)
1533 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1534
1535 if (db == NULL || db->db_level != 1 ||
1536 db->db_blkid >= end_blkid) {
1537 break;
1538 }
1539
1540 /*
1541 * Setup the next blkid we want to search for.
1542 */
1543 db_search.db_blkid = db->db_blkid + 1;
1544 ASSERT3U(db->db_blkid, >=, start_blkid);
1545
1546 /*
1547 * If the dbuf transitions to DB_EVICTING while we're trying
1548 * to dirty it, then we will be unable to discover it in
1549 * the dbuf hash table. This will result in a call to
1550 * dbuf_create() which needs to acquire the dn_dbufs_mtx
1551 * lock. To avoid a deadlock, we drop the lock before
1552 * dirtying the level-1 dbuf.
1553 */
1554 mutex_exit(&dn->dn_dbufs_mtx);
1555 dnode_dirty_l1(dn, db->db_blkid, tx);
1556 mutex_enter(&dn->dn_dbufs_mtx);
1557 }
1558
1559 #ifdef ZFS_DEBUG
1560 /*
1561 * Walk all the in-core level-1 dbufs and verify they have been dirtied.
1562 */
1563 db_search.db_level = 1;
1564 db_search.db_blkid = start_blkid + 1;
1565 db_search.db_state = DB_SEARCH;
1566 db = avl_find(&dn->dn_dbufs, &db_search, &where);
1567 if (db == NULL)
1568 db = avl_nearest(&dn->dn_dbufs, where, AVL_AFTER);
1569 for (; db != NULL; db = AVL_NEXT(&dn->dn_dbufs, db)) {
1570 if (db->db_level != 1 || db->db_blkid >= end_blkid)
1571 break;
1572 ASSERT(db->db_dirtycnt > 0);
1573 }
1574 #endif
1575 mutex_exit(&dn->dn_dbufs_mtx);
1576 }
1577
1578 void
dnode_free_range(dnode_t * dn,uint64_t off,uint64_t len,dmu_tx_t * tx)1579 dnode_free_range(dnode_t *dn, uint64_t off, uint64_t len, dmu_tx_t *tx)
1580 {
1581 dmu_buf_impl_t *db;
1582 uint64_t blkoff, blkid, nblks;
1583 int blksz, blkshift, head, tail;
1584 int trunc = FALSE;
1585 int epbs;
1586
1587 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1588 blksz = dn->dn_datablksz;
1589 blkshift = dn->dn_datablkshift;
1590 epbs = dn->dn_indblkshift - SPA_BLKPTRSHIFT;
1591
1592 if (len == DMU_OBJECT_END) {
1593 len = UINT64_MAX - off;
1594 trunc = TRUE;
1595 }
1596
1597 /*
1598 * First, block align the region to free:
1599 */
1600 if (ISP2(blksz)) {
1601 head = P2NPHASE(off, blksz);
1602 blkoff = P2PHASE(off, blksz);
1603 if ((off >> blkshift) > dn->dn_maxblkid)
1604 goto out;
1605 } else {
1606 ASSERT(dn->dn_maxblkid == 0);
1607 if (off == 0 && len >= blksz) {
1608 /*
1609 * Freeing the whole block; fast-track this request.
1610 */
1611 blkid = 0;
1612 nblks = 1;
1613 if (dn->dn_nlevels > 1)
1614 dnode_dirty_l1(dn, 0, tx);
1615 goto done;
1616 } else if (off >= blksz) {
1617 /* Freeing past end-of-data */
1618 goto out;
1619 } else {
1620 /* Freeing part of the block. */
1621 head = blksz - off;
1622 ASSERT3U(head, >, 0);
1623 }
1624 blkoff = off;
1625 }
1626 /* zero out any partial block data at the start of the range */
1627 if (head) {
1628 ASSERT3U(blkoff + head, ==, blksz);
1629 if (len < head)
1630 head = len;
1631 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off),
1632 TRUE, FALSE, FTAG, &db) == 0) {
1633 caddr_t data;
1634
1635 /* don't dirty if it isn't on disk and isn't dirty */
1636 if (db->db_last_dirty ||
1637 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1638 rw_exit(&dn->dn_struct_rwlock);
1639 dmu_buf_will_dirty(&db->db, tx);
1640 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1641 data = db->db.db_data;
1642 bzero(data + blkoff, head);
1643 }
1644 dbuf_rele(db, FTAG);
1645 }
1646 off += head;
1647 len -= head;
1648 }
1649
1650 /* If the range was less than one block, we're done */
1651 if (len == 0)
1652 goto out;
1653
1654 /* If the remaining range is past end of file, we're done */
1655 if ((off >> blkshift) > dn->dn_maxblkid)
1656 goto out;
1657
1658 ASSERT(ISP2(blksz));
1659 if (trunc)
1660 tail = 0;
1661 else
1662 tail = P2PHASE(len, blksz);
1663
1664 ASSERT0(P2PHASE(off, blksz));
1665 /* zero out any partial block data at the end of the range */
1666 if (tail) {
1667 if (len < tail)
1668 tail = len;
1669 if (dbuf_hold_impl(dn, 0, dbuf_whichblock(dn, 0, off+len),
1670 TRUE, FALSE, FTAG, &db) == 0) {
1671 /* don't dirty if not on disk and not dirty */
1672 if (db->db_last_dirty ||
1673 (db->db_blkptr && !BP_IS_HOLE(db->db_blkptr))) {
1674 rw_exit(&dn->dn_struct_rwlock);
1675 dmu_buf_will_dirty(&db->db, tx);
1676 rw_enter(&dn->dn_struct_rwlock, RW_WRITER);
1677 bzero(db->db.db_data, tail);
1678 }
1679 dbuf_rele(db, FTAG);
1680 }
1681 len -= tail;
1682 }
1683
1684 /* If the range did not include a full block, we are done */
1685 if (len == 0)
1686 goto out;
1687
1688 ASSERT(IS_P2ALIGNED(off, blksz));
1689 ASSERT(trunc || IS_P2ALIGNED(len, blksz));
1690 blkid = off >> blkshift;
1691 nblks = len >> blkshift;
1692 if (trunc)
1693 nblks += 1;
1694
1695 /*
1696 * Dirty all the indirect blocks in this range. Note that only
1697 * the first and last indirect blocks can actually be written
1698 * (if they were partially freed) -- they must be dirtied, even if
1699 * they do not exist on disk yet. The interior blocks will
1700 * be freed by free_children(), so they will not actually be written.
1701 * Even though these interior blocks will not be written, we
1702 * dirty them for two reasons:
1703 *
1704 * - It ensures that the indirect blocks remain in memory until
1705 * syncing context. (They have already been prefetched by
1706 * dmu_tx_hold_free(), so we don't have to worry about reading
1707 * them serially here.)
1708 *
1709 * - The dirty space accounting will put pressure on the txg sync
1710 * mechanism to begin syncing, and to delay transactions if there
1711 * is a large amount of freeing. Even though these indirect
1712 * blocks will not be written, we could need to write the same
1713 * amount of space if we copy the freed BPs into deadlists.
1714 */
1715 if (dn->dn_nlevels > 1) {
1716 uint64_t first, last;
1717
1718 first = blkid >> epbs;
1719 dnode_dirty_l1(dn, first, tx);
1720 if (trunc)
1721 last = dn->dn_maxblkid >> epbs;
1722 else
1723 last = (blkid + nblks - 1) >> epbs;
1724 if (last != first)
1725 dnode_dirty_l1(dn, last, tx);
1726
1727 dnode_dirty_l1range(dn, first, last, tx);
1728
1729 int shift = dn->dn_datablkshift + dn->dn_indblkshift -
1730 SPA_BLKPTRSHIFT;
1731 for (uint64_t i = first + 1; i < last; i++) {
1732 /*
1733 * Set i to the blockid of the next non-hole
1734 * level-1 indirect block at or after i. Note
1735 * that dnode_next_offset() operates in terms of
1736 * level-0-equivalent bytes.
1737 */
1738 uint64_t ibyte = i << shift;
1739 int err = dnode_next_offset(dn, DNODE_FIND_HAVELOCK,
1740 &ibyte, 2, 1, 0);
1741 i = ibyte >> shift;
1742 if (i >= last)
1743 break;
1744
1745 /*
1746 * Normally we should not see an error, either
1747 * from dnode_next_offset() or dbuf_hold_level()
1748 * (except for ESRCH from dnode_next_offset).
1749 * If there is an i/o error, then when we read
1750 * this block in syncing context, it will use
1751 * ZIO_FLAG_MUSTSUCCEED, and thus hang/panic according
1752 * to the "failmode" property. dnode_next_offset()
1753 * doesn't have a flag to indicate MUSTSUCCEED.
1754 */
1755 if (err != 0)
1756 break;
1757
1758 dnode_dirty_l1(dn, i, tx);
1759 }
1760 }
1761
1762 done:
1763 /*
1764 * Add this range to the dnode range list.
1765 * We will finish up this free operation in the syncing phase.
1766 */
1767 mutex_enter(&dn->dn_mtx);
1768 int txgoff = tx->tx_txg & TXG_MASK;
1769 if (dn->dn_free_ranges[txgoff] == NULL) {
1770 dn->dn_free_ranges[txgoff] = range_tree_create(NULL, NULL);
1771 }
1772 range_tree_clear(dn->dn_free_ranges[txgoff], blkid, nblks);
1773 range_tree_add(dn->dn_free_ranges[txgoff], blkid, nblks);
1774 dprintf_dnode(dn, "blkid=%llu nblks=%llu txg=%llu\n",
1775 blkid, nblks, tx->tx_txg);
1776 mutex_exit(&dn->dn_mtx);
1777
1778 dbuf_free_range(dn, blkid, blkid + nblks - 1, tx);
1779 dnode_setdirty(dn, tx);
1780 out:
1781
1782 rw_exit(&dn->dn_struct_rwlock);
1783 }
1784
1785 static boolean_t
dnode_spill_freed(dnode_t * dn)1786 dnode_spill_freed(dnode_t *dn)
1787 {
1788 int i;
1789
1790 mutex_enter(&dn->dn_mtx);
1791 for (i = 0; i < TXG_SIZE; i++) {
1792 if (dn->dn_rm_spillblk[i] == DN_KILL_SPILLBLK)
1793 break;
1794 }
1795 mutex_exit(&dn->dn_mtx);
1796 return (i < TXG_SIZE);
1797 }
1798
1799 /* return TRUE if this blkid was freed in a recent txg, or FALSE if it wasn't */
1800 uint64_t
dnode_block_freed(dnode_t * dn,uint64_t blkid)1801 dnode_block_freed(dnode_t *dn, uint64_t blkid)
1802 {
1803 void *dp = spa_get_dsl(dn->dn_objset->os_spa);
1804 int i;
1805
1806 if (blkid == DMU_BONUS_BLKID)
1807 return (FALSE);
1808
1809 /*
1810 * If we're in the process of opening the pool, dp will not be
1811 * set yet, but there shouldn't be anything dirty.
1812 */
1813 if (dp == NULL)
1814 return (FALSE);
1815
1816 if (dn->dn_free_txg)
1817 return (TRUE);
1818
1819 if (blkid == DMU_SPILL_BLKID)
1820 return (dnode_spill_freed(dn));
1821
1822 mutex_enter(&dn->dn_mtx);
1823 for (i = 0; i < TXG_SIZE; i++) {
1824 if (dn->dn_free_ranges[i] != NULL &&
1825 range_tree_contains(dn->dn_free_ranges[i], blkid, 1))
1826 break;
1827 }
1828 mutex_exit(&dn->dn_mtx);
1829 return (i < TXG_SIZE);
1830 }
1831
1832 /* call from syncing context when we actually write/free space for this dnode */
1833 void
dnode_diduse_space(dnode_t * dn,int64_t delta)1834 dnode_diduse_space(dnode_t *dn, int64_t delta)
1835 {
1836 uint64_t space;
1837 dprintf_dnode(dn, "dn=%p dnp=%p used=%llu delta=%lld\n",
1838 dn, dn->dn_phys,
1839 (u_longlong_t)dn->dn_phys->dn_used,
1840 (longlong_t)delta);
1841
1842 mutex_enter(&dn->dn_mtx);
1843 space = DN_USED_BYTES(dn->dn_phys);
1844 if (delta > 0) {
1845 ASSERT3U(space + delta, >=, space); /* no overflow */
1846 } else {
1847 ASSERT3U(space, >=, -delta); /* no underflow */
1848 }
1849 space += delta;
1850 if (spa_version(dn->dn_objset->os_spa) < SPA_VERSION_DNODE_BYTES) {
1851 ASSERT((dn->dn_phys->dn_flags & DNODE_FLAG_USED_BYTES) == 0);
1852 ASSERT0(P2PHASE(space, 1<<DEV_BSHIFT));
1853 dn->dn_phys->dn_used = space >> DEV_BSHIFT;
1854 } else {
1855 dn->dn_phys->dn_used = space;
1856 dn->dn_phys->dn_flags |= DNODE_FLAG_USED_BYTES;
1857 }
1858 mutex_exit(&dn->dn_mtx);
1859 }
1860
1861 /*
1862 * Scans a block at the indicated "level" looking for a hole or data,
1863 * depending on 'flags'.
1864 *
1865 * If level > 0, then we are scanning an indirect block looking at its
1866 * pointers. If level == 0, then we are looking at a block of dnodes.
1867 *
1868 * If we don't find what we are looking for in the block, we return ESRCH.
1869 * Otherwise, return with *offset pointing to the beginning (if searching
1870 * forwards) or end (if searching backwards) of the range covered by the
1871 * block pointer we matched on (or dnode).
1872 *
1873 * The basic search algorithm used below by dnode_next_offset() is to
1874 * use this function to search up the block tree (widen the search) until
1875 * we find something (i.e., we don't return ESRCH) and then search back
1876 * down the tree (narrow the search) until we reach our original search
1877 * level.
1878 */
1879 static int
dnode_next_offset_level(dnode_t * dn,int flags,uint64_t * offset,int lvl,uint64_t blkfill,uint64_t txg)1880 dnode_next_offset_level(dnode_t *dn, int flags, uint64_t *offset,
1881 int lvl, uint64_t blkfill, uint64_t txg)
1882 {
1883 dmu_buf_impl_t *db = NULL;
1884 void *data = NULL;
1885 uint64_t epbs = dn->dn_phys->dn_indblkshift - SPA_BLKPTRSHIFT;
1886 uint64_t epb = 1ULL << epbs;
1887 uint64_t minfill, maxfill;
1888 boolean_t hole;
1889 int i, inc, error, span;
1890
1891 dprintf("probing object %llu offset %llx level %d of %u\n",
1892 dn->dn_object, *offset, lvl, dn->dn_phys->dn_nlevels);
1893
1894 hole = ((flags & DNODE_FIND_HOLE) != 0);
1895 inc = (flags & DNODE_FIND_BACKWARDS) ? -1 : 1;
1896 ASSERT(txg == 0 || !hole);
1897
1898 if (lvl == dn->dn_phys->dn_nlevels) {
1899 error = 0;
1900 epb = dn->dn_phys->dn_nblkptr;
1901 data = dn->dn_phys->dn_blkptr;
1902 } else {
1903 uint64_t blkid = dbuf_whichblock(dn, lvl, *offset);
1904 error = dbuf_hold_impl(dn, lvl, blkid, TRUE, FALSE, FTAG, &db);
1905 if (error) {
1906 if (error != ENOENT)
1907 return (error);
1908 if (hole)
1909 return (0);
1910 /*
1911 * This can only happen when we are searching up
1912 * the block tree for data. We don't really need to
1913 * adjust the offset, as we will just end up looking
1914 * at the pointer to this block in its parent, and its
1915 * going to be unallocated, so we will skip over it.
1916 */
1917 return (SET_ERROR(ESRCH));
1918 }
1919 error = dbuf_read(db, NULL, DB_RF_CANFAIL | DB_RF_HAVESTRUCT);
1920 if (error) {
1921 dbuf_rele(db, FTAG);
1922 return (error);
1923 }
1924 data = db->db.db_data;
1925 }
1926
1927
1928 if (db != NULL && txg != 0 && (db->db_blkptr == NULL ||
1929 db->db_blkptr->blk_birth <= txg ||
1930 BP_IS_HOLE(db->db_blkptr))) {
1931 /*
1932 * This can only happen when we are searching up the tree
1933 * and these conditions mean that we need to keep climbing.
1934 */
1935 error = SET_ERROR(ESRCH);
1936 } else if (lvl == 0) {
1937 dnode_phys_t *dnp = data;
1938 span = DNODE_SHIFT;
1939 ASSERT(dn->dn_type == DMU_OT_DNODE);
1940
1941 for (i = (*offset >> span) & (blkfill - 1);
1942 i >= 0 && i < blkfill; i += inc) {
1943 if ((dnp[i].dn_type == DMU_OT_NONE) == hole)
1944 break;
1945 *offset += (1ULL << span) * inc;
1946 }
1947 if (i < 0 || i == blkfill)
1948 error = SET_ERROR(ESRCH);
1949 } else {
1950 blkptr_t *bp = data;
1951 uint64_t start = *offset;
1952 span = (lvl - 1) * epbs + dn->dn_datablkshift;
1953 minfill = 0;
1954 maxfill = blkfill << ((lvl - 1) * epbs);
1955
1956 if (hole)
1957 maxfill--;
1958 else
1959 minfill++;
1960
1961 *offset = *offset >> span;
1962 for (i = BF64_GET(*offset, 0, epbs);
1963 i >= 0 && i < epb; i += inc) {
1964 if (BP_GET_FILL(&bp[i]) >= minfill &&
1965 BP_GET_FILL(&bp[i]) <= maxfill &&
1966 (hole || bp[i].blk_birth > txg))
1967 break;
1968 if (inc > 0 || *offset > 0)
1969 *offset += inc;
1970 }
1971 *offset = *offset << span;
1972 if (inc < 0) {
1973 /* traversing backwards; position offset at the end */
1974 ASSERT3U(*offset, <=, start);
1975 *offset = MIN(*offset + (1ULL << span) - 1, start);
1976 } else if (*offset < start) {
1977 *offset = start;
1978 }
1979 if (i < 0 || i >= epb)
1980 error = SET_ERROR(ESRCH);
1981 }
1982
1983 if (db)
1984 dbuf_rele(db, FTAG);
1985
1986 return (error);
1987 }
1988
1989 /*
1990 * Find the next hole, data, or sparse region at or after *offset.
1991 * The value 'blkfill' tells us how many items we expect to find
1992 * in an L0 data block; this value is 1 for normal objects,
1993 * DNODES_PER_BLOCK for the meta dnode, and some fraction of
1994 * DNODES_PER_BLOCK when searching for sparse regions thereof.
1995 *
1996 * Examples:
1997 *
1998 * dnode_next_offset(dn, flags, offset, 1, 1, 0);
1999 * Finds the next/previous hole/data in a file.
2000 * Used in dmu_offset_next().
2001 *
2002 * dnode_next_offset(mdn, flags, offset, 0, DNODES_PER_BLOCK, txg);
2003 * Finds the next free/allocated dnode an objset's meta-dnode.
2004 * Only finds objects that have new contents since txg (ie.
2005 * bonus buffer changes and content removal are ignored).
2006 * Used in dmu_object_next().
2007 *
2008 * dnode_next_offset(mdn, DNODE_FIND_HOLE, offset, 2, DNODES_PER_BLOCK >> 2, 0);
2009 * Finds the next L2 meta-dnode bp that's at most 1/4 full.
2010 * Used in dmu_object_alloc().
2011 */
2012 int
dnode_next_offset(dnode_t * dn,int flags,uint64_t * offset,int minlvl,uint64_t blkfill,uint64_t txg)2013 dnode_next_offset(dnode_t *dn, int flags, uint64_t *offset,
2014 int minlvl, uint64_t blkfill, uint64_t txg)
2015 {
2016 uint64_t initial_offset = *offset;
2017 int lvl, maxlvl;
2018 int error = 0;
2019
2020 if (!(flags & DNODE_FIND_HAVELOCK))
2021 rw_enter(&dn->dn_struct_rwlock, RW_READER);
2022
2023 if (dn->dn_phys->dn_nlevels == 0) {
2024 error = SET_ERROR(ESRCH);
2025 goto out;
2026 }
2027
2028 if (dn->dn_datablkshift == 0) {
2029 if (*offset < dn->dn_datablksz) {
2030 if (flags & DNODE_FIND_HOLE)
2031 *offset = dn->dn_datablksz;
2032 } else {
2033 error = SET_ERROR(ESRCH);
2034 }
2035 goto out;
2036 }
2037
2038 maxlvl = dn->dn_phys->dn_nlevels;
2039
2040 for (lvl = minlvl; lvl <= maxlvl; lvl++) {
2041 error = dnode_next_offset_level(dn,
2042 flags, offset, lvl, blkfill, txg);
2043 if (error != ESRCH)
2044 break;
2045 }
2046
2047 while (error == 0 && --lvl >= minlvl) {
2048 error = dnode_next_offset_level(dn,
2049 flags, offset, lvl, blkfill, txg);
2050 }
2051
2052 /*
2053 * There's always a "virtual hole" at the end of the object, even
2054 * if all BP's which physically exist are non-holes.
2055 */
2056 if ((flags & DNODE_FIND_HOLE) && error == ESRCH && txg == 0 &&
2057 minlvl == 1 && blkfill == 1 && !(flags & DNODE_FIND_BACKWARDS)) {
2058 error = 0;
2059 }
2060
2061 if (error == 0 && (flags & DNODE_FIND_BACKWARDS ?
2062 initial_offset < *offset : initial_offset > *offset))
2063 error = SET_ERROR(ESRCH);
2064 out:
2065 if (!(flags & DNODE_FIND_HAVELOCK))
2066 rw_exit(&dn->dn_struct_rwlock);
2067
2068 return (error);
2069 }
2070